Introduction to Networking and Systems Measurements Lecture 2: Basic Network Measurements Dr Noa Zilberman [email protected] Prof Andrew W. Moore [email protected] Networking and Systems Measurements(L50) 1 Terminology Matters! … in greater depth in following weeks Networking and Systems Measurements (L50) 2 Bandwidth, Throughput and Goodput § Bandwidth – how much data can pass through a channel. § Throughput – how much data actually travels through a channel. § Goodput is often referred to as application level throughput. But bandwidth can be limited below link’s capacity and vary over time, throughput can be measured differently from bandwidth etc….. Networking and Systems Measurements (L50) 3 Speed and Bandwidth § Higher bandwidth does not necessarily mean higher speed § E.g. can mean the aggregation of links Ø 100G = 2x50G or 4x25G or 10x10G Ø A very common practice in interconnects Networking and Systems Measurements (L50) 4 RTT, Latency and FCT Measures of time: § Latency – The time interval between two events. § Round Trip Time (RTT) – The time interval between a signal being transmitted and a reply is being received. § Flow Completion Time (FCT) – The lifetime of a flow. Networking and Systems Measurements (L50) 5 Performance Metrics § Throughput, FCT etc. are measures of Performance. § Bandwidth, RTT, packet loss etc. don’t indicate (directly) how good or bad the application / system / network perform. Networking and Systems Measurements (L50) 6 Example: The Effect of Latency on Application’s Performance Networking and Systems Measurements (L50) 7 Example: The Effect of Latency on Application’s Performance Memcached Server performance drops to 60% with the addition of 200μs of additional latency Networking and Systems Measurements (L50) 8 Types of Measurements Networking and Systems Measurements (L50) 9 Measurement Techniques § Active ØIssue probe, Analyse response § Passive ØObserve events Networking and Systems Measurements(L50) 10 Example: Active vs. Passive RTT Measurement § Active measurement – Ping Ø Sends ICMP Echo Request message Ø Waits for Echo Reply message Ø RTT is the time gap between the request and the reply. § Passive measurement – tcptrace Ø Uses TCP dump files Ø Calculates RTT according to timestamps logged in the dump. Networking and Systems Measurements(L50) 11 Comparison Passive Active Can only measure in the presence Measures even when tapping of activity / traffic activity / traffic is not possible Measures user experience, Measures system, network, behaviour application performance Measures protocol exchanges Raise privacy concerns Adds probing load: - Overload system/network - May bias inferences Networking and Systems Measurements(L50) 12 Measurement Vantage Point § Point where measurement host connects to system / network § Observations often depend on vantage point Ø Do you have enough vantage points? Ø How are the vantage points distributed? § Can affect, e.g.: Ø Topology discovery Ø Bandwidth analysis Networking and Systems Measurements (L50) 13 Possible Vantage Points § End-hosts Ø Active measurements of end-to-end paths Ø Passive measurements of host’s traffic § Routers/Measurement hosts in network Ø Active measurements of network paths Ø Passive measurements of traffic, protocol exchanges, configuration Networking and Systems Measurements (L50) 14 Time flies § 1ns = 20cm in fibre § 10Gb/s is about 10 bits per nanosecond § so at 10Gb/s a 512byte packet is ~ 8meters long Networking and Systems Measurements (L50) 15 Ping § Ping is basically a “are you still there” test § “connectivity” test § “how long does it take to get there” test § “loss approximation” test $ ping www.stanford.edu PING www.stanford.edu (54.192.2.121): 56 data bytes 64 bytes from 54.192.2.121: icmp_seq=0 ttl=242 time=3.730 ms 64 bytes from 54.192.2.121: icmp_seq=1 ttl=242 time=3.845 ms … ^C --- www.stanford.edu ping statistics --- 8 packets transmitted, 8 packets received, 0.0% packet loss round-trip min/avg/max/stddev = 3.730/3.808/3.849/0.047 ms Networking and Systems Measurements (L50) 16 PING traps § Uses ICMP (control messages of the Internet) § Might not follow the same path as normal packets § Might be filtered § A ping test is not the actual round trip time for an application § One way delay is not simply twice round trip time § Learn by doing (run tcpdump at the same time) Networking and Systems Measurements (L50) 17 Recall the Internet federation § The Internet ties together different networks Ø >18,000 ISP networks user ISP A ISP B ISP C user We can see (hints) of the nodes and links using traceroute… 18 Traceroute: Internet debug thy self § Recall the Internet Zombie plan – Time-To-Live (TTL) § Each router decrements TTL; when TTL =0 send error Traceroute artificially sets low TTL and receives the error Each step of the path is iteratively discovered TTL=1 0 TTL=0 so “error” from first step TTL=2 1 0 TTL=0 so “error” from second step TTL=3 2 1 0 TTL=0 so “error” from third step TTL=4 3 2 1 0 TTL=0 so “error” from forth step … Networking and Systems Measurements (L50) 19 Traceroute as we wish… But ONLY one direction Networking and Systems Measurements (L50) 20 “Real” Internet traceroute traceroute: rio.cl.cam.ac.uk to munnari.oz.au (tracepath on windows is similar) Three delay measurements from traceroute munnari.oz.au rio.cl.cam.ac.uk to gatwick.net.cl.cam.ac.uk traceroute to munnari.oz.au (202.29.151.3), 30 hops max, 60 byte packets 1 gatwick.net.cl.cam.ac.uk (128.232.32.2) 0.416 ms 0.384 ms 0.427 ms 2 cl-sby.route-nwest.net.cam.ac.uk (193.60.89.9) 0.393 ms 0.440 ms 0.494 ms trans-continent 3 route-nwest.route-mill.net.cam.ac.uk (192.84.5.137) 0.407 ms 0.448 ms 0.501 ms link 4 route-mill.route-enet.net.cam.ac.uk (192.84.5.94) 1.006 ms 1.091 ms 1.163 ms 5 xe-11-3-0.camb-rbr1.eastern.ja.net (146.97.130.1) 0.300 ms 0.313 ms 0.350 ms 6 ae24.lowdss-sbr1.ja.net (146.97.37.185) 2.679 ms 2.664 ms 2.712 ms 7 ae28.londhx-sbr1.ja.net (146.97.33.17) 5.955 ms 5.953 ms 5.901 ms 8 janet.mx1.lon.uk.geant.net (62.40.124.197) 6.059 ms 6.066 ms 6.052 ms 9 ae0.mx1.par.fr.geant.net (62.40.98.77) 11.742 ms 11.779 ms 11.724 ms 10 ae1.mx1.mad.es.geant.net (62.40.98.64) 27.751 ms 27.734 ms 27.704 ms 11 mb-so-02-v4.bb.tein3.net (202.179.249.117) 138.296 ms 138.314 ms 138.282 ms 12 sg-so-04-v4.bb.tein3.net (202.179.249.53) 196.303 ms 196.293 ms 196.264 ms 13 th-pr-v4.bb.tein3.net (202.179.249.66) 225.153 ms 225.178 ms 225.196 ms 14 pyt-thairen-to-02-bdr-pyt.uni.net.th (202.29.12.10) 225.163 ms 223.343 ms 223.363 ms 15 202.28.227.126 (202.28.227.126) 241.038 ms 240.941 ms 240.834 ms 16 202.28.221.46 (202.28.221.46) 287.252 ms 287.306 ms 287.282 ms 17 * * * 18 * * * * means no response (probe lost, router not replying) 19 * * * 20 coe-gw.psu.ac.th (202.29.149.70) 241.681 ms 241.715 ms 241.680 ms 21 munnari.OZ.AU (202.29.151.3) 241.610 ms 241.636 ms 241.537 ms 21 Traceroute traps – a bit like ping § Uses UDP or ICMP (but traffic is often TCP) § Might not follow the same path § Might be filtered § Only infers one direction of the path § Replies can be very weird § One way delay is not simply twice round trip time (networks may have many paths) § Learn by doing (try with and without the –I option) Networking and Systems Measurements (L50) 22 Traceroute doesn't always know Networking and Systems Measurements (L50) 23 Traceroute lies Networking and Systems Measurements (L50) 24 Beyond traceroute § Paris traceroute Uses many proBes to identify multiple paths www.paris-traceroute.net § Reverse traceroute Uses a remote server to proBe reverse path Networking and Systems Measurements (L50) 25 Link capacity….. § Recall capacity is a property of where and what we measure § Nominal network capacity is physical eg 100BaseTX Ethernet: 100 Mbps WiFi 802.11g: 54 Mbps § IP-layer capacity < nominal capacity Ø Coding schemes Ø Framing bits, overhead Ø Medium access control Networking and Systems Measurements (L50) 26 Link capacity & utilization § Link capacity (C(Δt)) ≈ IP-layer capacity Ø Maximum IP-layer rate of maximum-sized packets Ø IP-layer capacity depends on sizeLink of utilizationpacket relative to layer- 2 overhead • Link utilization (u(Δt)) § Link utilization (u(Δt)) ! u(Δt) = Average bits transmitted on the link during Δt Ø u(Δt) = Average bits transmitted on! Percent the link utilization during = Δt% link capacity that is utilized Ø Percent utilization = capacity % link capacity that is utilized utilization bitsper second time 3 Networking and Systems Measurements (L50) 27 Link Availableavailable bandwidth Capacity • Available bandwidth (A(Δt)) § Available bandwidth (A(Δt)) Ø Maximum! Maximum unused bandwidthunused bandwidth Ø A(Δt)! =A( C(ΔΔtt)) =- u(C(ΔtΔ) t) - u(Δt) capacity available bandwidth utilization bitsper second time 4 Networking and Systems Measurements (L50) 28 End-to-end capacity and End-to-end effective bandwidth Router1 -----C1----- Router2 -----C2----- Router C1: 100 Mbps C2: 30 Mbps u1: 80 Mbps u2: 3 Mbps A1: 20 Mbps A2: 27 Mbps End-to-end capacity: min{C1, C2}=30 Mbps End-to-end available bandwidth: min{A1, A2}=20 Mbps Networking and Systems Measurements (L50) 29 Probing method Flooding Advanced methods Issue enough probes to “fill” A number of methods in path literature: § Pro Packet pair, size-delay, self- Ø Measure what users can get induced congestion § Con § Pro Ø Large overhead affects Ø Less overhead than flooding network and users § Con Ø Rely on assumptions that don’t always hold in practice Networking and Systems Measurements (L50) 30 Effective Bandwidth Measurement § How much capacity in my network? Ø Is it working at spec.? Am I getting my money’s worth? Ø Systems can adapt to change of Effective Bandwidth Networking and Systems Measurements (L50) 31 Considerations § TCP versus UDP Ø UDP not biased by congestion/flow control Ø Flooding with UDP may create too much congestion and bias results Ø Multiple TCP connections reduces bias § Multi-threaded TCP Ø How many threads? Ø Which size transfers? § UDP Ø How to pick sending rate? Networking and Systems Measurements (L50) 32 iperf versions and other tools for measuring available bandwidth § iperf/iperf3 Ø Control of client and server Ø Configurable tests § iperf2 for UDP § iperf3 is a rewrite with different/improved TCP Others: eg.